A supervised machine learning-based framework to detect low-level fault injections in software systems

Abstract

Fault injection attacks inject faults into system components, inducing abnormal software behavior. Software vulnerability analysis cannot prevent new attack vectors without software modifications. Attack detection methods utilize system-specific software features and unsupervised learning due to lack of labelled data. Unsupervised pattern recognition is vulnerable to false data injection, and Machine Learning algorithms such as Artificial and Recurrent Neural Networks are not feasible for resource-constrained software systems. Supervised detection of low-level attack effects presents a possible solution to these issues. This thesis introduces a supervised ML-based framework to detect low-level software fault injections consisting of labelled dataset generation using an instruction-level software fault injection tool to simulate attack effects. The framework is implemented on two software systems and the results demonstrate its feasibility. The thesis explores system-level threat detection due to simulated low-level attack effects and demonstrates that combining application data and software properties improves the low-level software fault injection prediction.